In one week, GPT-5.6 proved a 50-year-old conjecture, closed a 30-year-old optimization gap, and solved every IMO 2026 problem, signaling that LLMs have crossed a threshold in formal mathematical reasoning

The Week AI Proved Real Theorems

For fifty years, the Cycle Double Cover Conjecture resisted every mathematician who attempted it. The problem, posed independently in the 1970s by Paul Seymour, George Szekeres, and W. T. Tutte, asks whether every bridgeless graph contains a collection of cycles that covers every edge exactly twice. One sentence to state. Fifty years unsolved.

A language model proved it in under an hour.

OpenAI announced that GPT-5.6 Sol Ultra produced a proof of the conjecture using 64 sub-agents, completing the work in under one hour 1. The announcement was meant as a showcase. It became a manual.

Days later, a teaching professor in UC Berkeley's Industrial Engineering and Operations Research department, who holds a PhD in applied mathematics, adapted the same prompting methodology to a problem closer to home. The target was a complexity gap in zeroth-order convex optimization that had stood open since 1996, a class of problems where the optimizer works without access to gradient information. In a single 148-minute session, GPT-5.6 Sol Pro produced a proof that closed the gap. The result was formally verified in Lean, the open-source proof assistant that checks mathematical arguments with the same mechanical rigor a compiler applies to source code 2.

The researcher disclosed that the optimization result has not yet been peer reviewed 2. In ordinary mathematical practice, that caveat would stall the conversation. Lean works differently. It does not convene panels or weigh reputations. It checks each logical inference, step by step, against a kernel of axioms. If the proof compiles, it is correct. If it does not compile, it is wrong. No referee fatigue, no halo effect from a prestigious byline, no argument that sounds airtight until someone finds the gap on page forty.

That distinction separates this week from years of inflated claims about AI and mathematics. Language models have been generating proof-like text since well before GPT-5. The bottleneck was never generation. It was trust. You could not rely on the output without a human mathematician checking every line, and human mathematicians are finite in number and hours. Pairing a language model with Lean collapses that bottleneck. The model proposes. The proof assistant verifies. The check is mechanical, deterministic, and complete.

One more result surfaced the same week. A user documented GPT-5.6 solving all six problems from the 2026 International Mathematical Olympiad. It does not carry the Lean verification that makes the CDC and optimization proofs airtight, but it sketches a trajectory.

One result is an anecdote. Two is a data point. Three, across graph theory, convex optimization, and competition mathematics, in the same week, is a pattern. Large language models, given structured prompts and paired with formal verification tools, are producing novel proofs that survive expert scrutiny. The common ingredient is not any single model or any single problem. It is the combination of capable generation and mechanical checking.

For researchers, the practical signal is already visible. The UC Berkeley professor did not engineer a new system. He took an existing prompting methodology from someone else's announcement, pointed it at a problem in his own domain, and closed a gap that had resisted three decades of human effort in two and a half hours. That transferability is what separates a lab demo from infrastructure. When a methodology works well enough that a single researcher can replicate it on an unrelated open problem in an afternoon, it has crossed from novelty into utility.

The question for the research community is no longer whether language models can produce correct proofs. Lean answers that. The question is which open problems fall next, and how many mathematicians will spend the coming year finding out.